📄 In-Depth Analysis of Arabic-Origin Words in the Turkish Morpholex

Summary

The Turkish Morpholex is a manually built morphological lexicon of 48,472 words taken from the KeNet wordnet. It treats Arabic loanwords as ordinary Turkish stems. For example, it stops at adalet when segmenting adaletli, so it misses the real Arabic root. The authors fill this gap. They used the etymology labels of the official TDK dictionary to find Arabic-origin entries in the lexicon. Then they annotated each one by hand following Arabic grammar: its consonantal root, its morphological pattern (wazn) and its semantic/grammatical category. The main reference was the Riyadh Dictionary, with the Doha Dictionary as a backup. The annotators faced three kinds of hard cases. Some words are obsolete in modern Arabic. Some are Turkish coinages built on Arabic patterns. Others are compounds of several Arabic elements. The result is an open dataset that links thousands of Turkish words to about 1,400 Arabic roots, and annotator agreement was high. The resource makes Turkish morphological analysis more accurate, helps NLP handle loanwords, and offers a workflow that could transfer to Urdu and Persian.

Key points

• The starting lexicon is the Turkish Morpholex (Arıcan et al., 2022): 48,472 words drawn from the KeNet wordnet and segmented into roots, prefixes and suffixes.
• The TDK digital dictionary yielded 4,687 unique Arabic-origin words. Some TDK labels turned out to be wrong (e.g., Patlıcan, Sabun). These were excluded and the errors were documented.
• The final mapping links 3,855 source Turkish roots to 1,430 distinct Arabic roots.
• Three annotators did the work, all native Arabic speakers fluent in Turkish. They resolved disagreements in weekly sessions and reached 93% agreement with the consensus.
• The most frequent semantic category is the "meaning noun" (1,789 words). Common patterns include Taf‘īl, Fa‘l, Fā‘il, Maf‘ūl, If‘āl and Ift‘āl. The annotations are released on GitHub (turkish-lexicon-arabic-roots).

Interesting ideas

• I1. Borrowed words should be segmented by the morphology of the donor language rather than forced into native stem-plus-suffix analysis, otherwise the lexicon stops at the wrong root.
• I2. An external authority (TDK etymology) is used only to propose candidates, and expert review then audits it, keeping a documented list of wrong labels.
• I3. Hard cases get explicit labels instead of being forced into the standard scheme: obsolete-in-source words, hybrid coinages and multi-root compounds each carry notes.
• I4. Annotating root, pattern and semantic category together exposes derivational families, such as mahkeme, hikmet and hakem sharing one root.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): The English lemmatizer could gain a donor-aware layer for Latinate and Greek forms (irregular plurals, bound prefixes and stems). This would give deeper and more correct lemma and derivation analyses than surface suffix stripping.
• I2 (learn): External resources and LLM teachers can be treated as fallible witnesses whose suggestions pass a verification step. Every rejected item and its reason goes into a log, which can later serve as negative controls.
• I3 (tokens): Morphology entries could carry flags such as "hybrid", "compound" or "no clean root". These would let downstream steps know when an analysis is uncertain, keeping each step explainable.
• I4 (world): The compiled knowledge layer could link derivational relatives, such as a verb with its agent noun and its nominalization. Then a frame found through one word form could be connected to entities and events named by another.